MadryLab / trak

A fast, effective data attribution method for neural networks in PyTorch
https://trak.csail.mit.edu/
MIT License
180 stars 26 forks source link

RuntimeError: CUDA error: the provided PTX was compiled with an unsupported toolchain. #68

Closed simplelifetime closed 6 months ago

simplelifetime commented 6 months ago

What might be the caused of this problem, example code: import fast_jl

fast_jl.project_rademacher_8(torch.zeros( 8, 1_000, device=device), 512, 0, num_sms)

cuda version: Cuda compilation tools, release 11.8, V11.8.89

nvidia driver: 460.91.03

kristian-georgiev commented 6 months ago

This is likely due to a mismatch between your nvidia driver and cuda toolkit version, and not related to TRAK. I would start from the NVIDIA docs: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/#running-the-binaries.

simplelifetime commented 6 months ago

Thanks, updating driver to 530.35 and cuda version to 12.1 solve the problem

shangqing-liu commented 5 months ago

@simplelifetime Hi I also met this problem which the exact driver version can you provide to solve this problem. Thanks.